Automated Endocardial Border Detection and Left Ventricular Functional Assessment in Echocardiography Using Deep...

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate endocardial border detection is crucial for assessing left ventricular systolic function in echocardiography.
• Current manual methods are time-consuming and prone to observer variability.
• Automated, accurate detection methods are needed to improve clinical workflow.

Purpose of the Study:

• To develop a deep learning-based method for automated endocardial border detection and left ventricular functional assessment in echocardiographic videos.
• To compare the performance of four deep learning segmentation models: U-Net, UNet++, UNet3+, and Deep Residual U-Net.

Main Methods:

• Segmentation of the left ventricular cavity was performed using U-Net, UNet++, UNet3+, and Deep Residual U-Net in 2D echocardiographic videos.
• Performance was evaluated using intersection over union and Dice coefficient metrics.
• Accuracy was assessed by calculating the mean estimation error for left ventricular ejection fraction, global longitudinal strain, and global circumferential strain.

Main Results:

• UNet++ and UNet3+ demonstrated high performance in segmentation accuracy (intersection over union and Dice coefficient).
• UNet++ outperformed other methods with acceptable mean estimation errors: 10.8% for ejection fraction, 8.5% for global longitudinal strain, and 5.8% for global circumferential strain.
• The developed UNet++ method showed the best overall performance.

Conclusions:

• The deep learning-based method using UNet++ offers a promising solution for automated endocardial border detection in echocardiography.
• This approach has the potential to enhance accuracy, reduce variability, and improve the efficiency of left ventricular functional assessment.
• The method may serve as a valuable tool for echocardiography examiners, optimizing clinical workflows.